It looks like there is a way to turn off the detection for vehicle or human so to be fair I doubt many would be looking for autos in a small hallway scene. However, that error leads me to believe they may be using the size of an object in the scene to differentiate between human or auto?
For sure, if I set up a camera in that hallway in production, I would likely not have checked the "vehicle" detection box. That's exactly the problem. If we had checked only person detection in that scene, we would have missed a significant number of subjects. They simply aren't detected at all.
It also leads me to believe that they are not doing any kind of true object classification, most commonly done today with CNN/DNN analytics. Kind of curious given their use of the Movidius chip, from this test they do not seem to be fully utilizing its capabilities.
We've added a section comparing Acusense to Hikvision's standard Intrusion and VMD:
Vs. Hikvision Standard Intrusion And VMD
Compared to Hikvision's standard Intrusion analytics, Acusense performance is clearly better, rejecting more false alarms from more sources, including shadows and lights, more alarms on animals, etc. Performance is drastically better than standard VMD, which alerted on practically any pixel change, including but not limited to the common false alert sources above, without discriminating between people, vehicles, or other objects.
Interestingly enough, the Hik website I primarily use does not even show a 5.5.72 firmware. It shows the current one now as 5.6.3, and the only previous version shown is 5.6.0. Ditto for the other Accusense models I checked. (5.6.3 is also available for many other cameras without Accusense.)
...get an alarm while I was running, even having it at max wouldn't trigger an alarm, it did not even draw a bounding box around me.
SoC: Did you see that?
AI: See what?
SoC: If I knew “what”, I wouldn’t need you. Something just went flying by...
AI: You mean that blurred streak of light. I have no idea. It’s gone now. Anyway, out of frame, out of mind, right?
SoC: Well I’m gonna send an alert anyway, the user should be notified!
AI: And what are you going to say it is this time, a giraffe, a car, maybe a sand storm?
SoC: Nah, you’re right those are getting stale. You got any ideas?
AI: Yeah, keep your trap shut. Remember, if faced with the choice of false negative vs. false positive, ALWAYS choose false negative. If you missed it, they probably will too; nobody’s the wiser. On the other hand, calling out “heavy snow” in July will just put us on the cover of IPVM, again!
There is one thing I do not understand. Analytics got settings, right? And those probably need to be tuned for different conditions. Are you sure you do a good job on tunning it? Or do you always use out of the box ones?
For clarity, we've added two sections to the report. As in other tests, we reached out to Hikvision for comment and suggestions on optimizing the analytics. We also perform our own settings changes to see what effects they have (in this case, minimal to none).
Note that Hikvision was given the opportunity to review results during testing but did not offer any suggestions for settings changes to improve detection. They pointed out that Acusense cameras worked best in conjunction with Acusense NVRs, but in our tests we saw no difference in performance between Acusense cameras' on board detection and Acusense cameras used in conjunction with the NVR.
Settings Changes Tested In All Scenes
In all scenes, we adjusted settings to see if detection performance improved, but found that default settings produced the best results. In Hikvision's intrusion analytics, including Acusense, the only settings adjustments are threshold (time object is in scene before triggering) and sensitivity ("size of the object which can trigger alarm"). Neither of these settings eliminated false alerts.
My understanding of Hik's Acusense is you can get these analytics on-board the camera as well as built into the NVR. The NVR will work the analytics on ordinary camera feeds, while Acusense cameras will provide the analytics on NON-Acusense NVRs by just pushing through an event notification.
Is this correct? If so I'm thinking of picking up one of the Acusense NVRs to replace the one we have at the shop which has various IP cameras with the basic analytics. Thoughts?
Yes, UI3, this is correct, the Acusense NVR analytics will work with third party camera streams and provide people and vehicle alerts. The Acusense cameras will have event notifications on non-Acusense NVRs but won't provide the classification.
Is it still the case that standard motion events (VMD) is disabled when the AcuSense analytics are enabled resulting in continuous recording? I tested the DS-2CD2T46G1 when it first came out and could not have both VMD and AcuSense enabled simultaneously.